doctr.transforms

Data transformations are part of both training and inference procedure. Drawing inspiration from the design of torchvision, we express transformations as composable modules.

Supported transformations

Here are all transformations that are available through docTR:

class doctr.transforms.Resize(output_size: int | Tuple[int, int], method: str = 'bilinear', preserve_aspect_ratio: bool = False, symmetric_pad: bool = False)

Resizes a tensor to a target size

>>> import tensorflow as tf
>>> from doctr.transforms import Resize
>>> transfo = Resize((32, 32))
>>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Args:

output_size: expected output size method: interpolation method preserve_aspect_ratio: if True, preserve aspect ratio and pad the rest with zeros symmetric_pad: if True while preserving aspect ratio, the padding will be done symmetrically

class doctr.transforms.Normalize(mean: Tuple[float, float, float], std: Tuple[float, float, float])

Normalize a tensor to a Gaussian distribution for each channel

>>> import tensorflow as tf
>>> from doctr.transforms import Normalize
>>> transfo = Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
>>> out = transfo(tf.random.uniform(shape=[8, 64, 64, 3], minval=0, maxval=1))

Args:

mean: average value per channel std: standard deviation per channel

class doctr.transforms.LambdaTransformation(fn: Callable[[Tensor], Tensor])

Normalize a tensor to a Gaussian distribution for each channel

>>> import tensorflow as tf
>>> from doctr.transforms import LambdaTransformation
>>> transfo = LambdaTransformation(lambda x: x/ 255.)
>>> out = transfo(tf.random.uniform(shape=[8, 64, 64, 3], minval=0, maxval=1))

Args:

fn: the function to be applied to the input tensor

class doctr.transforms.ToGray(num_output_channels: int = 1)

Convert a RGB tensor (batch of images or image) to a 3-channels grayscale tensor

>>> import tensorflow as tf
>>> from doctr.transforms import ToGray
>>> transfo = ToGray()
>>> out = transfo(tf.random.uniform(shape=[8, 64, 64, 3], minval=0, maxval=1))

class doctr.transforms.ColorInversion(min_val: float = 0.5)

Applies the following tranformation to a tensor (image or batch of images): convert to grayscale, colorize (shift 0-values randomly), and then invert colors

TensorFlowPyTorch

>>> import tensorflow as tf
>>> from doctr.transforms import ColorInversion
>>> transfo = ColorInversion(min_val=0.6)
>>> out = transfo(tf.random.uniform(shape=[8, 64, 64, 3], minval=0, maxval=1))

Args:

min_val: range [min_val, 1] to colorize RGB pixels

class doctr.transforms.RandomBrightness(max_delta: float = 0.3)

Randomly adjust brightness of a tensor (batch of images or image) by adding a delta to all pixels

>>> import tensorflow as tf
>>> from doctr.transforms import RandomBrightness
>>> transfo = RandomBrightness()
>>> out = transfo(tf.random.uniform(shape=[8, 64, 64, 3], minval=0, maxval=1))

Args:

max_delta: offset to add to each pixel is randomly picked in [-max_delta, max_delta] p: probability to apply transformation

class doctr.transforms.RandomContrast(delta: float = 0.3)

Randomly adjust contrast of a tensor (batch of images or image) by adjusting each pixel: (img - mean) * contrast_factor + mean.

>>> import tensorflow as tf
>>> from doctr.transforms import RandomContrast
>>> transfo = RandomContrast()
>>> out = transfo(tf.random.uniform(shape=[8, 64, 64, 3], minval=0, maxval=1))

Args:

delta: multiplicative factor is picked in [1-delta, 1+delta] (reduce contrast if factor<1)

class doctr.transforms.RandomSaturation(delta: float = 0.5)

Randomly adjust saturation of a tensor (batch of images or image) by converting to HSV and increasing saturation by a factor.

>>> import tensorflow as tf
>>> from doctr.transforms import RandomSaturation
>>> transfo = RandomSaturation()
>>> out = transfo(tf.random.uniform(shape=[8, 64, 64, 3], minval=0, maxval=1))

Args:

delta: multiplicative factor is picked in [1-delta, 1+delta] (reduce saturation if factor<1)

class doctr.transforms.RandomHue(max_delta: float = 0.3)

Randomly adjust hue of a tensor (batch of images or image) by converting to HSV and adding a delta

>>> import tensorflow as tf
>>> from doctr.transforms import RandomHue
>>> transfo = RandomHue()
>>> out = transfo(tf.random.uniform(shape=[8, 64, 64, 3], minval=0, maxval=1))

Args:

max_delta: offset to add to each pixel is randomly picked in [-max_delta, max_delta]

class doctr.transforms.RandomGamma(min_gamma: float = 0.5, max_gamma: float = 1.5, min_gain: float = 0.8, max_gain: float = 1.2)

randomly performs gamma correction for a tensor (batch of images or image)

>>> import tensorflow as tf
>>> from doctr.transforms import RandomGamma
>>> transfo = RandomGamma()
>>> out = transfo(tf.random.uniform(shape=[8, 64, 64, 3], minval=0, maxval=1))

Args:

min_gamma: non-negative real number, lower bound for gamma param max_gamma: non-negative real number, upper bound for gamma min_gain: lower bound for constant multiplier max_gain: upper bound for constant multiplier

class doctr.transforms.RandomJpegQuality(min_quality: int = 60, max_quality: int = 100)

Randomly adjust jpeg quality of a 3 dimensional RGB image

>>> import tensorflow as tf
>>> from doctr.transforms import RandomJpegQuality
>>> transfo = RandomJpegQuality()
>>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Args:

min_quality: int between [0, 100] max_quality: int between [0, 100]

class doctr.transforms.RandomRotate(max_angle: float = 5.0, expand: bool = False)

Randomly rotate a tensor image and its boxes

Args:

max_angle: maximum angle for rotation, in degrees. Angles will be uniformly picked in

[-max_angle, max_angle]

expand: whether the image should be padded before the rotation

class doctr.transforms.RandomCrop(scale: Tuple[float, float] = (0.08, 1.0), ratio: Tuple[float, float] = (0.75, 1.33))

Randomly crop a tensor image and its boxes

Args:

scale: tuple of floats, relative (min_area, max_area) of the crop ratio: tuple of float, relative (min_ratio, max_ratio) where ratio = h/w

class doctr.transforms.GaussianBlur(kernel_shape: int | Iterable[int], std: Tuple[float, float])

Randomly adjust jpeg quality of a 3 dimensional RGB image

>>> import tensorflow as tf
>>> from doctr.transforms import GaussianBlur
>>> transfo = GaussianBlur(3, (.1, 5))
>>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Args:

kernel_shape: size of the blurring kernel std: min and max value of the standard deviation

class doctr.transforms.ChannelShuffle

Randomly shuffle channel order of a given image

class doctr.transforms.GaussianNoise(mean: float = 0.0, std: float = 1.0)

Adds Gaussian Noise to the input tensor

>>> import tensorflow as tf
>>> from doctr.transforms import GaussianNoise
>>> transfo = GaussianNoise(0., 1.)
>>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Args:

mean : mean of the gaussian distribution std : std of the gaussian distribution

class doctr.transforms.RandomHorizontalFlip(p: float)

Adds random horizontal flip to the input tensor/np.ndarray

>>> import tensorflow as tf
>>> from doctr.transforms import RandomHorizontalFlip
>>> transfo = RandomHorizontalFlip(p=0.5)
>>> image = tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1)
>>> target = np.array([[0.1, 0.1, 0.4, 0.5] ], dtype= np.float32)
>>> out = transfo(image, target)

Args:

p : probability of Horizontal Flip

class doctr.transforms.RandomShadow(opacity_range: Tuple[float, float] | None = None)

Adds random shade to the input image

>>> import tensorflow as tf
>>> from doctr.transforms import RandomShadow
>>> transfo = RandomShadow(0., 1.)
>>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Args:

opacity_range : minimum and maximum opacity of the shade

class doctr.transforms.RandomResize(scale_range: Tuple[float, float] = (0.3, 0.9), preserve_aspect_ratio: bool | float = False, symmetric_pad: bool | float = False, p: float = 0.5)

Randomly resize the input image and align corresponding targets

>>> import tensorflow as tf
>>> from doctr.transforms import RandomResize
>>> transfo = RandomResize((0.3, 0.9), preserve_aspect_ratio=True, symmetric_pad=True, p=0.5)
>>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Args:

scale_range: range of the resizing factor for width and height (independently) preserve_aspect_ratio: whether to preserve the aspect ratio of the image,

given a float value, the aspect ratio will be preserved with this probability

symmetric_pad: whether to symmetrically pad the image,

given a float value, the symmetric padding will be applied with this probability

p: probability to apply the transformation

Composing transformations

It is common to require several transformations to be performed consecutively.

class doctr.transforms.Compose(transforms: List[Callable[[Any], Any]])

Implements a wrapper that will apply transformations sequentially

>>> import tensorflow as tf
>>> from doctr.transforms import Compose, Resize
>>> transfos = Compose([Resize((32, 32))])
>>> out = transfos(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Args:

transforms: list of transformation modules

class doctr.transforms.OneOf(transforms: List[Callable[[Any], Any]])

Randomly apply one of the input transformations

TensorFlowPyTorch

>>> import tensorflow as tf
>>> from doctr.transforms import OneOf
>>> transfo = OneOf([JpegQuality(), Gamma()])
>>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Args:

transforms: list of transformations, one only will be picked

class doctr.transforms.RandomApply(transform: Callable[[Any], Any], p: float = 0.5)

Apply with a probability p the input transformation

TensorFlowPyTorch

>>> import tensorflow as tf
>>> from doctr.transforms import RandomApply
>>> transfo = RandomApply(Gamma(), p=.5)
>>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))

Args:

transform: transformation to apply p: probability to apply